Multi-Aspect Full-System Server Model and Optimization Concept As a Simulation-Based Approach (MFSMOS)
Total Page:16
File Type:pdf, Size:1020Kb
Multi-aspect Full-system Server Model and Optimization Concept as a Simulation-based Approach (MFSMOS) Dissertation in Computer Science A thesis submitted to the Faculty of Computer Science, Electrical Engineering and Mathematics, Paderborn University in partial fulfillment of the requirements for the degree of doctor rerum naturalium (Dr. rer. nat.) by Diana Riemer Paderborn, Germany, 2017-11-29 Supervisors: Prof. Dr.-Ing. André Brinkmann Johannes Gutenberg University Mainz Prof. Dr. Franz Josef Rammig Paderborn University ii Abstract The exponential growth of the information technology, such as social media content or on- demand video services, results in increasing energy consumption, which may be one of the main causes of climate changes. Power and cooling are the key challenges to reducing the emission of greenhouse gas, especially in large-scale data centers. With our full-system model, we contribute towards improving the energy efficiency on specific server systems and, consequently, optimize the efficiency of data center infrastructures. We develop a generic, flexible, and scalable model to simulate and optimize a complete server system for the multiple, potentially conflicting aspects: power, temperature, and performance. We develop a hierarchical and abstract model of a rack-mounted server system, which builds the base of our mathematical methods to calculate the multi-aspects of each component. We demonstrate the feasibility and advantages of our concept through a prototypical implementation, in which we empirically validate our model using a variety of artificial workloads to ensure the reproducibility at any time. In principle, our simulation-based, full- system server model supports customer-specific workload scenarios, specified as realistic category-specific utilization levels, to simulate the suitable power of server systems. We address the significant static as well as dynamic characteristics and configurations to cover a variety of server systems and components compatible with the customer-specific server system. We precisely calculate the power consumption that reduces the over-provisioning of the server system, particularly in industrial practice. Moreover, we demonstrate that we can forecast future generations of high-performance systems and components by assuming the predecessor or a similar generation. To our knowledge, in academic research, there are no generic approaches that cover the full server system simulation on a common base. This thesis provides new research contributions that explicitly cover the heterogeneous characteristics of the hardware and software variations, such as supporting diverse server families or generations. Moreover, the simulation optimizes the energy efficiency of the server system at various utilization levels, especially at low-intensity phases (under- utilization). iii Zusammenfassung Der ansteigende Medienkonsum, bedingt durch sich weiterentwickelnde Kommunikations- kanäle, führt zum kontinuierlichen Anstieg des Energiebedarfs, welches den Klimawandel weiter voranschreiten lässt. Das größte Potenzial zur Energieeinsparung im Rechenzentrum wird derzeit dem IT-Equipment sowie dessen Kühlung zugesprochen. Wir haben ein ganzheitliches Systemmodell zur Optimierung der Energieeffizienz von Rack-Servern entwickelt. Wir entwickelten ein allgemein-gültiges, flexibles und skalierbares Model, um verschiedene Server zu simulieren und optimieren. Das hierarchische und abstrahierte Servermodell unterstützt die Berechnung der konkurrierenden Aspekte: Energieverbrauch, Temperaturentwicklung und Performance. Die prototypische Implementierung zeigt die Machbarkeit und Vorteile unseres Ansatzes. In der Evaluation unseres simulationsbasierten Systemmodells verwenden wir synthetische Lastszenarien zur besseren Nachvollziehbarkeit, wobei wir auch realistische benutzerspezifische Lastszenarien unterstützen. Der Anwender definiert die prozentuale Auslastung der Komponenten als Lastszenario, welches wir zur Berechnung der maximalen Leistungsaufnahme und des Energieverbrauchs verwenden. Wir präsentieren die relevanten statischen und dynamischen Merkmale, um unterschiedliche Serversysteme und Komponenten aus verschiedenen Generationen abzubilden. Wir berechnen den Energieverbrauch bzw. die Leistungsaufnahme und die daraus resultierenden Temperaturen hinreichend genau, welches die oftmals erhebliche Überschätzung des tatsächlichen Energieverbrauchs von Servern reduziert. Unser Ansatz ermöglicht die Prognose der maximalen Leistungsaufnahme von zukünftigen Systemen und Komponenten. Unser Ansatz unterstützt, im Gegensatz zu den bisherigen akademischen Ansätzen, eine Vielzahl an Server und deren unterschiedliche Komponenten. Mithilfe unseres simulationsbasierten Ansatzes können wir die Energieeffizienz von Servern bei jeglichen anwenderspezifischen Szenarien optimieren. iv Acknowledgement First and foremost, I would like to express my deep appreciation to Professor Dr.-Ing. André Brinkmann and Professor Dr. Franz J. Rammig, my research supervisors, for their patient guidance, vital support, immense knowledge, and constructive suggestions during the planning and development of this dissertation. I could never have imagined having better mentors for my PhD study. In particular, I sincerely thank Franz for his extreme patience in the face of numerous obstacles and his invaluable advice towards the success of my thesis. Besides my advisors, I would like to thank my committee members, Professor Dr. Christian Plessl, Professor Dr. Gudrun Oevel, and Dr. Wolfgang Müller for their time, insightful comments, and hard questions in the examination board. I would like to pay special thanks to Cornelia Wiederhold who was periodically busy with printing, binding and delivering this thesis. I would also like to express my very profound gratitude to Dr. Wolfgang Müller and Bernhard Homölle, who initiated the cooperation between C-LAB at Paderborn University and Fujitsu Technology Solutions GmbH. My regards go out to the management of both persons for the third-party project and for the contributions, that each of them made to my intellectual growth during my years of study. I would like to thank the following staffers at Fujitsu Technology Solutions GmbH as well as Atos IT Solutions and Services GmbH: Karsten Beins, Hansfried Block, Dr. Martin Clemens, Thomas Diekmann, Helmut Emmerich, Wolfgang Erig, Stefan Festl, Robert Hoffmann, Sebastian Keßler, Brigitte Krawinkel, Alois Lichtenstern, Johannes Linne, Bernhard Lüthen, Karl-Josef Lüttgenau, Uwe Meiners, Georg Müller, Dr. Oliver Niehörster, Detlef Pusch, Klaus Rammig, Dr. Mitsuru Sato, Bernhard Schräder, Dr. Jürgen Schrage, Jothiram Selvam, Dieter Tenberg, Sylvia Tuschen, Walter Unruh, Bernd Winkelsträter, and Norbert Zobe. You provided enormous support and were always willing to help me. I must express my very great appreciation to Harald Lüdtke for providing me valuable support with Kalcheck, immense knowledge, and continuous motivation throughout the ongoing process of researching and writing this thesis. I would also like to extend my thanks to the technicians of the environmental laboratory and hardware department at Fujitsu Technology Solutions GmbH for their technical support, such as collecting the server- specific data, providing the necessary IT equipment, and introducing me to the tools for my measurements and analysis. This dissertation was partially supported by the German Research Foundation (DFG) within the Collaborative Research Centre On-The-Fly Computing (SFB901). My special thanks are extended to Dr. Michelle R. Kloppenburg, who assisted through proofreading, my English teacher Mary Cremer, and classmates Annika Ballhausen, Michaela Kemper, Kirstin Köhler, Birgit Petermeier, and Annette Zaloudek for enhancing my speaking language skills. v Furthermore, I would like to show my gratitude to my fellow doctoral students Peer Adelt, Markus Becker, Dr. Gilles Bertrand Gnokam Defo, Jan Jatzkowski, Mabel Mary Joy, Dr. Alexander Jungmann, Dr. Matthias Keller, Christoph Kuznik, Fabian Mischkalla, Sven Schönberg, Dr. Katharina Stahl, and Jan Stenner for their feedback, conversations, and of course friendship. I would like to offer my special thanks to Dr. Lorijn van Rooijen for her guidance in defining precise mathematical formulas. Getting through my studies required more than academic support and I would like to offer my special thanks to Jan Braun, Claudia Dobrinski, Marlies Drestomark, Saikal Gerlach, Isabel Junger, Andreas Kaufmann, Dr. Isabel Koke, Christoph Nagel, and Dr. Gaby Nordendorf for listening to me, tolerating me, and supporting me over the last several years. Indeed, it would have been a lonely office without Veit Dornseifer, Mohamed Hassan Mohamed Elsharkawi, Bastian Koppelmann, Arjun Ramaswami, and Jonathan Westerholt. Finally, I especially thank Ginger Claassen for our daily conversations and your morning-based reliable coffee delivery. I would also like to thank my friends for their wise counsel and infinite support. Furthermore, I am particularly grateful for my family and, especially, my parents who raised me with a tremendous passion for technology, particularly in computer science, and supported me in the pursuit of this project. They were always supporting me and encouraging me with their best wishes. I would like to thank Bernd Rummler very much for his widespread support, giving me confidence, and always being there for me. Finally, I would like to thank all those mentors, advisors and friends who have supported me and believed